Building Governed ERP Agents on AWS: Series Guide
A technical series on building governed AI agents for ERP exception handling on AWS: identity, agentic SOPs, policy guardrails, and observability.
Series: Building Production-Grade ERP Agents on AWS (18 articles)
- 1Building Governed ERP Agents on AWS: Series Guide This article
Building governed ERP agents on AWS: series guide
Month-end leaves a pile of ERP exceptions that all need judgment: invoices that do not match their purchase order, goods receipts that never arrived, accruals nobody can explain. An agent can work them, and the hard part turns out not to be the model. It is identity, procedure, policy, and proof. That is what this series is about, and every post answers one question completely so you can read in order or jump straight to the one you need.
Series: Agentic ERP Automation Quickstart | builder.aws.com
Start here
The AWS Agentic Standard Operating Procedures (SOPs) for ERP Automation post explains the SOP-driven automation pattern. This technical series explains how to use and extend the Agentic ERP Automation Quickstart , an AWS-maintained open source reference implementation for ERP exception workflows on Amazon Bedrock AgentCore.
This series excludes customer-specific outcomes and named customer stories until communications review confirms their public-release status.
What this series covers
- Deploy the reference implementation in your own AWS account.
- Choose an SAP identity path that fits your environment.
- Design, author, and operate agentic SOPs.
- Understand the controls that apply to your deployed topology.
- Extend the implementation with exception types and SAP APIs.
- Interpret observability evidence before operating at scale.
Before you begin
- An AWS account with Amazon Bedrock model access enabled
- Node.js 20 or later, Python 3.10 or later, and configured AWS CLI credentials
- SAP is optional for the initial deployment. The application runs without a SAP connection.
- Model inference cost depends on your own exception mix, procedure length, and model choice. The first post quotes the benchmark figures recorded in the repository, from one measured run on one set of test cases. Measure your own deployment with an AWS Cost Explorer report before you plan against a number.
The series
Phase 1: Get running
| Post | What you get |
|---|---|
| Deploy an AI agent for SAP exception handling | Deploy, log in, and verify a running agent in your own account. SAP is optional. |
| How an agent authenticates to SAP | One setting, four decisions, and how to read what each profile is verified to do. |
Phase 2: Auth setup
| Post | What you get |
|---|---|
cognito-basic: a technical SAP user | The verified shared-technical-user path. It does not give you per-user SAP attribution. |
cognito-m2m: SAP client credentials | The deployable client-credentials path, and what you have to verify yourself. |
entra-obo: per-user SAP identity | The verified attended per-user path. It reaches SAP directly, not through a Gateway target. |
Phase 3: The SOP model
| Post | What you get |
|---|---|
| What is an agentic SOP? | The line between a skill and a procedure, and why procedures are configuration. |
| How to write an SOP an agent can follow | Structure, constants, entity sets, and getting your procedure into the knowledge base. |
| How an agent loads its procedure | The three runtime paths, and what each one substitutes into the prompt. |
Phase 4: Governance and trust
| Post | What you get |
|---|---|
| Cedar policy for AI agents | The three conditions that all have to hold before a policy blocks a request. |
| Agent autonomy and unattended mode | What trigger-mode starts, and the write control people mistake it for. |
Phase 5: Extend and operate
| Post | What you get |
|---|---|
| Adding a new exception type | Tell a configuration edit from a much larger change before you start. |
| Connect a custom SAP OData service | The application-side workflow, and the SAP-side work nothing can accelerate for you. |
| Single agent vs multi-agent | When an orchestrator and specialist split earns the extra operational boundary. |
| AI agent observability | Which signal answers which question, and the four limits that catch operators out. |
Phase 6: Architecture explainers
| Post | What you get |
|---|---|
| Which agent changes need a deployment | Which changes publish as content, and which need a deployment. |
| AgentCore Gateway vs a self-hosted MCP server | Why Gateway fronts some SAP paths and not others. |
| S3 Vectors vs OpenSearch Serverless | Why this sample uses Amazon S3 Vectors, and when it would not. |
Public resources
| Asset | Link |
|---|---|
| Reference implementation | Agentic ERP Automation Quickstart |
| AWS for SAP MCP Server | Documentation |
| Amazon Bedrock AgentCore | Product page |
| Strands Agents SDK | GitHub repository |
| ABAP Accelerator | AWS for SAP blog |
Series: Building Production-Grade ERP Agents on AWS (18 articles)
- 1Building Governed ERP Agents on AWS: Series Guide This article
Enjoyed reading this content? Let the author know!
Your likes, comments, shares, and saves help creators reach more builders.
Loading recommendations
Loading article